akm-cli 0.9.0-rc.0 → 0.9.0-rc.13

package/dist/workflows/exec/step-work.js

@@ -0,0 +1,1190 @@
+// This Source Code Form is subject to the terms of the Mozilla Public
+// License, v. 2.0. If a copy of the MPL was not distributed with this
+// file, You can obtain one at https://mozilla.org/MPL/2.0/.
+/**
+ * Shared step semantics — the ONE implementation of a step's orchestration
+ * decisions, consumed by BOTH the engine loop (`run-workflow.ts` +
+ * `native-executor.ts`) and, from R3 on, the harness-neutral driver protocol
+ * (`workflow brief` / `workflow report`). The cardinal rule of the driver
+ * protocol (redesign addendum R3) is *no duplicated semantics*: work-list
+ * computation, prompt assembly, reducer/artifact promotion, output-schema
+ * validation, artifact-judged gate summaries, gate-feedback recovery, and
+ * route evaluation live here so an engine-driven run and a brief/report-driven
+ * run of the same frozen plan produce byte-identical unit graphs.
+ *
+ * ## What is PURE here
+ *
+ * {@link computeStepWorkList} — given the frozen step plan and a
+ * {@link WorkListInput} (params, prior step outputs, gate-loop number + its
+ * recovered feedback) — is a pure function: same inputs ⇒ same unit ids, input
+ * hashes, and fully-resolved prompts. It takes NO clock, NO IO, and NO journal
+ * (journal-derived state, i.e. the recovered gate feedback, is passed in). This
+ * is the load-bearing guarantee that `brief` can predict exactly the units the
+ * engine would dispatch. So are the reducer/artifact helpers
+ * ({@link buildEvidence}, {@link projectStepOutput}, {@link validateStepArtifact},
+ * {@link buildArtifactSummary}), the gate-feedback recovery
+ * ({@link recoverGateFeedback} / {@link activeGateLoop}), and route evaluation
+ * ({@link evaluateRoute} and its bookkeeping).
+ *
+ * ## What does IO here
+ *
+ * The gate-evaluation journaling ({@link journalGateEvaluationStart} /
+ * {@link journalGateEvaluationFinish}) writes `workflow_run_units` rows through
+ * the serialized writer queue — an engine-driven judge call is an LLM call and
+ * is journaled like a unit. It lives here (not in the engine loop) so the
+ * report path journals gate evaluations through the identical writer.
+ *
+ * This module NEVER dispatches a unit and NEVER writes step rows: dispatch is
+ * the executor's job (`native-executor.ts`), advancing the gated spine is the
+ * engine loop's job (`run-workflow.ts` via `completeWorkflowStep`).
+ */
+import { createHash } from "node:crypto";
+import unitPreambleTemplate from "../../assets/prompts/workflow-unit-preamble.md" with { type: "text" };
+import { UsageError } from "../../core/errors.js";
+import { appendEvent } from "../../core/events.js";
+import { validateJsonSchemaSubset } from "../../core/json-schema.js";
+import { withWorkflowRunsRepo } from "../../storage/repositories/workflow-runs-repository.js";
+import { canonicalJson as canonicalJsonString } from "../ir/plan-hash.js";
+import { resolveReferenceString } from "../program/expressions.js";
+import { WORKFLOW_MAX_MAP_EXPANSION } from "../resource-limits.js";
+import { requireExecutableWorkflowPlan } from "../runtime/plan-classifier.js";
+import { completeWorkflowStep } from "../runtime/runs.js";
+import { GATE_EVALUATION_PHASE } from "../runtime/unit-phases.js";
+import { enqueueUnitWrite } from "./unit-writer.js";
+/**
+ * Default per-unit timeout. Deliberately NOT the 60 s agent default
+ * (`DEFAULT_AGENT_TIMEOUT_MS`) — workflow units routinely run real coding
+ * tasks on slow local models; 10 minutes matches the LLM-path default
+ * (`tryLlmFeature`). A unit's `timeout` declaration overrides this; `none`
+ * disables.
+ */
+export const DEFAULT_UNIT_TIMEOUT_MS = 600_000;
+/** How much raw unit output is retained in step evidence (full text lives on the unit row). */
+const EVIDENCE_TEXT_CLIP = 2_000;
+/** How much artifact JSON the completion-criteria judge receives (addendum R2, artifact-judging gates). */
+const GATE_ARTIFACT_CLIP = 4_000;
+/**
+ * Compute a step's expected work-list PURELY from the frozen plan and its
+ * inputs: resolve the fan-out list, derive content-derived unit ids, assemble
+ * each unit's prompt (preamble + interpolated instructions + gate feedback +
+ * schema directive), and hash the resolved input. Same inputs ⇒ byte-identical
+ * ids/hashes/prompts — the invariant `brief` relies on to predict the engine.
+ *
+ * Whole-list failures (missing subgraph, unresolvable / non-array `over`,
+ * null or duplicate fan-out items) return `{ ok: false }`. The per-unit
+ * `resolved: { ok: false }` branch is STRUCTURALLY UNREACHABLE in the unified
+ * format — prose is never scanned for references, and everything that CAN
+ * fail (map.over / route.input / inputs:) resolves once per step, failing the
+ * whole list above. The branch is retained because brief/report/executor all
+ * share the shape and defensively handle it; if a future unit kind
+ * reintroduces per-unit resolution (e.g. an exec/shell unit with real
+ * substitution), the cross-surface failure plumbing is already in place.
+ */
+/**
+ * Validate a fan-out item list BEFORE any identity/dispatch work: expansion
+ * within the resource limit, no null/undefined items, no canonical duplicates.
+ * Returns the failure message, or undefined when the list is dispatchable.
+ *
+ * Null items: producer garbage — there is nothing to hand the unit as its work
+ * item. The pre-unification format rejected them incidentally (substituting
+ * `${{ item }}` failed); with items attached as context instead of spliced,
+ * nothing later would stop a unit from being dispatched with "Item: null", so
+ * the rejection is explicit here. Duplicates: content-derived unit identity
+ * makes canonical duplicates collide on id — an authoring error caught
+ * deterministically, before dispatch.
+ */
+function validateFanOutItems(stepId, items) {
+    if (items.length > WORKFLOW_MAX_MAP_EXPANSION) {
+        return `Step "${stepId}" fan-out expands to ${items.length} units, exceeding the ${WORKFLOW_MAX_MAP_EXPANSION}-unit resource limit.`;
+    }
+    const nullIndex = items.findIndex((item) => item === null || item === undefined);
+    if (nullIndex !== -1) {
+        return (`Step "${stepId}" fan-out list contains a null item (index ${nullIndex}). ` +
+            `Every item must be a concrete value — fix the producing step's output.`);
+    }
+    const firstIndexByCanonical = new Map();
+    for (let i = 0; i < items.length; i++) {
+        const canonical = canonicalJson(items[i]) ?? "null";
+        const firstIndex = firstIndexByCanonical.get(canonical);
+        if (firstIndex !== undefined) {
+            return (`Step "${stepId}" fan-out list contains duplicate items (indices ${firstIndex} and ${i}: ` +
+                `${clip(canonical, 200)}). Content-derived unit identity requires distinct items — ` +
+                `deduplicate the list this workflow fans out over.`);
+        }
+        firstIndexByCanonical.set(canonical, i);
+    }
+    return undefined;
+}
+export function computeStepWorkList(plan, input) {
+    const root = plan.root;
+    // Route-only steps (YAML `route:`) carry no execution subgraph.
+    if (!root) {
+        return {
+            ok: false,
+            error: `Step "${plan.stepId}" has no execution subgraph (a route-only step); the native executor cannot dispatch it.`,
+        };
+    }
+    const template = root.kind === "map" ? root.template : root;
+    const reducer = root.kind === "map" ? root.reducer : "collect";
+    const scope = { params: input.params, stepOutputs: input.stepOutputs };
+    // Instructions are ALWAYS the step's body prose, byte-exact — never
+    // templated, never scanned for reference syntax (workflow-format-
+    // unification, spec §2.3). Only `map.over` / `route.input` / `inputs[]`
+    // carry the closed reference grammar.
+    // Resolve the step's declared `inputs:` ONCE (shared by every unit in this
+    // step — map items differ, declared inputs do not): prior-step artifacts
+    // attached to every dispatched unit as structured context.
+    const resolvedInputs = [];
+    for (const reference of template.inputs ?? []) {
+        const resolved = resolveReferenceString(reference, scope);
+        if (!resolved.ok) {
+            return {
+                ok: false,
+                error: `Step "${plan.stepId}" declared input "${reference}" failed to resolve: ${resolved.error.message}`,
+            };
+        }
+        resolvedInputs.push({ reference, value: resolved.value });
+    }
+    // Resolve fan-out items: `over` is a single whole-value reference naming
+    // its producer explicitly — no ambient key search.
+    let items;
+    if (root.kind === "map") {
+        const source = resolveReferenceString(root.over, scope);
+        if (!source.ok) {
+            return {
+                ok: false,
+                error: `Step "${plan.stepId}" fan-out "over" (${root.over}) failed to resolve: ${source.error.message}`,
+            };
+        }
+        if (!Array.isArray(source.value)) {
+            return {
+                ok: false,
+                error: `Step "${plan.stepId}" fan-out "over" (${root.over}) resolved to ${typeof source.value}, not an array.`,
+            };
+        }
+        items = source.value;
+    }
+    else {
+        items = [undefined];
+    }
+    const isFanOut = root.kind === "map";
+    const fanOutProblem = isFanOut ? validateFanOutItems(plan.stepId, items) : undefined;
+    if (fanOutProblem)
+        return { ok: false, error: fanOutProblem };
+    // Content-derived unit identity: compute every id up front (duplicate items
+    // were rejected above — identity requires distinct items).
+    const unitIds = items.map((item) => unitIdFor(template.id, item, isFanOut));
+    const gateLoop = input.gateLoop ?? 1;
+    const frozenInvocation = template.invocation;
+    if (!frozenInvocation)
+        return { ok: false, error: `Step "${plan.stepId}" has no frozen invocation.` };
+    const frozenEngine = input.engines?.[frozenInvocation.engine];
+    if (!frozenEngine) {
+        return { ok: false, error: `Step "${plan.stepId}" references missing frozen engine "${frozenInvocation.engine}".` };
+    }
+    const runner = frozenEngine.kind === "llm" ? "llm" : frozenEngine.runnerKind;
+    const timeoutMs = frozenInvocation.timeoutMs;
+    const units = items.map((item, index) => {
+        const unitId = unitIds[index];
+        // Gate loops (>= 2) journal under `<unitId>~l<loop>` so loop 1's rows are
+        // never clobbered; the content-derived identity (and the prompt's
+        // {{UNIT_ID}}) stays the base id.
+        const journalBaseId = gateLoop > 1 ? `${unitId}~l${gateLoop}` : unitId;
+        // Context attachment (workflow-format-unification, spec §4): every unit
+        // receives the run params (already in the preamble), its item + index if
+        // it is a map unit, and the artifacts named by its step's `inputs:`.
+        // Instructions reach the unit byte-exact — never interpolated.
+        const prompt = buildUnitPrompt({
+            runId: input.runId,
+            stepId: plan.stepId,
+            unitId,
+            params: input.params,
+            ...(isFanOut ? { item, itemIndex: index } : {}),
+            ...(resolvedInputs.length > 0 ? { inputs: resolvedInputs } : {}),
+            ...(input.gateFeedback ? { gateFeedback: input.gateFeedback } : {}),
+            ...(template.schema ? { schema: template.schema } : {}),
+            instructions: template.instructions,
+        });
+        // Canonical dispatch-input envelope (reviewer finding #1). Every field
+        // here is a PLAN-FROZEN input that changes what the backend is actually
+        // asked to do, so a completed unit is reused ONLY when all of them match;
+        // a change to any of them re-dispatches. Key order is FIXED — it is the
+        // hash preimage (JSON.stringify preserves insertion order) — and shared
+        // by ALL surfaces, since this is the ONE place a unit's inputHash is
+        // computed (engine, brief, and report all call computeStepWorkList), so
+        // the byte-identical hash across surfaces is structural, not coincidental.
+        //
+        // Unit identity (workflow-format-unification, spec §2.3/§4) hashes the
+        // FROZEN TEMPLATE BYTES (`template.instructions`, byte-exact, never an
+        // instantiated/interpolated string) + the canonical item JSON + the
+        // declared-input artifact hashes + the params snapshot — instead of a
+        // resolved/spliced prompt string, since there is no more splicing. The
+        // assembled `prompt` above is what the harness SEES; the hash is over the
+        // plan-frozen INPUTS that determine it, which is the same replay contract
+        // the old resolved-prompt hash gave (same inputs ⇒ same hash) with the
+        // interpolation step removed.
+        //
+        // Included beyond the R4 baseline (template/runner/model/schema): resolved
+        // timeoutMs, the env asset ref NAMES, and isolation — each reaches
+        // dispatch (native-executor's UnitDispatchRequest) and a changed one
+        // yields a materially different call. `env` carries NAMES ONLY, never
+        // resolved values: hashing a resolved secret would leak it into a durable
+        // hash oracle and would spuriously re-dispatch on every secret rotation.
+        // `retry`/`onError` are DELIBERATELY excluded — they govern failed-unit
+        // re-dispatch and step-level failure reduction, not a COMPLETED unit's
+        // inputs/output, so a completed row stays valid across policy changes.
+        //
+        // `gateFeedback` IS included (conditionally, so a no-feedback unit's
+        // preimage is byte-identical to before): it is appended to the prompt by
+        // `buildUnitPrompt`, so a gate loop's retry is materially a different ask
+        // than the rejected attempt — omitting it made loop 1 and loop 2 journal
+        // identical hashes for different prompts, breaking the "changed inputs ⇒
+        // changed hash" audit contract. Replay-safe: feedback is re-derived from
+        // the journaled gate decision, so a resumed retry re-hashes identically.
+        //
+        // Ambient config is DELIBERATELY excluded — the model-alias table, the
+        // resolved backend/connection, and the working directory (`ctx.workDir` /
+        // process.cwd()) are NOT plan-frozen. The frozen plan is the identity
+        // boundary (redesign addendum determinism bar #2): config drift under an
+        // in-flight run is out of scope by design.
+        const dispatch = transitiveDispatchSnapshot(frozenEngine, input.engines ?? {});
+        const inputHash = createHash("sha256")
+            .update(canonicalJsonString({
+            hashVersion: 4,
+            template: template.instructions,
+            item: isFanOut ? (item ?? null) : null,
+            inputs: resolvedInputs,
+            params: input.params,
+            dispatch,
+            invocation: frozenInvocation,
+            schema: template.schema ?? null,
+            env: template.env ?? null,
+            isolation: template.isolation ?? "none",
+            ...(input.gateFeedback ? { gateFeedback: input.gateFeedback } : {}),
+        }))
+            .digest("hex");
+        const resolved = { ok: true, prompt, inputHash };
+        return {
+            unitId,
+            nodeId: template.id,
+            index,
+            item,
+            isFanOut,
+            journalBaseId,
+            runner,
+            engine: frozenEngine,
+            ...(frozenEngine?.kind === "agent" &&
+                frozenEngine.fallbackLlmEngine &&
+                input.engines?.[frozenEngine.fallbackLlmEngine]?.kind === "llm"
+                ? {
+                    fallbackEngine: input.engines[frozenEngine.fallbackLlmEngine],
+                }
+                : {}),
+            invocation: frozenInvocation,
+            ...(frozenInvocation.model ? { model: frozenInvocation.model } : {}),
+            timeoutMs,
+            ...(template.schema ? { schema: template.schema } : {}),
+            ...(template.env ? { env: template.env } : {}),
+            ...(template.retry ? { retry: template.retry } : {}),
+            onError: template.onError,
+            ...(template.isolation ? { isolation: template.isolation } : {}),
+            resolved,
+        };
+    });
+    const concurrency = root.kind === "map" ? root.concurrency : 1;
+    return {
+        ok: true,
+        list: { template, reducer, isFanOut, ...(concurrency !== undefined ? { concurrency } : {}), items, units },
+    };
+}
+/**
+ * Assemble the final prompt: engine preamble (run params + item/index +
+ * declared-input artifacts, all as structured JSON context) + the step's
+ * BYTE-EXACT prose instructions (+ gate feedback on loop re-executions, +
+ * schema directive). Instructions are NEVER interpolated (workflow-format-
+ * unification, spec §2.3) — data reaches the unit as attached context, not
+ * string splices; only the ENGINE's own preamble placeholders are substituted
+ * here.
+ */
+export function buildUnitPrompt(input) {
+    const { runId, stepId, unitId, params, itemIndex, item, inputs, gateFeedback, schema, instructions } = input;
+    // Function replacements throughout: a string replacement would interpret
+    // GetSubstitution patterns ($&, $$, $', $`) inside VALUES and silently
+    // corrupt the prompt (e.g. a param value containing "$&").
+    const preamble = unitPreambleTemplate
+        .replaceAll("{{RUN_ID}}", () => runId)
+        .replaceAll("{{STEP_ID}}", () => stepId)
+        .replaceAll("{{UNIT_ID}}", () => unitId)
+        .replaceAll("{{PARAMS_JSON}}", () => safeJson(params));
+    // Map-unit context: the item this unit was given, plus its index. Attached
+    // as structured JSON — the engine never splices it into the instructions.
+    const itemBlock = itemIndex !== undefined
+        ? `\n\n## Item (index ${itemIndex})\nYou were given this item from the fan-out list:\n${safeJson(item)}`
+        : "";
+    // Declared `inputs:` context: the prior-step artifacts this step named.
+    const inputsBlock = inputs && inputs.length > 0
+        ? `\n\n## Declared inputs\n${inputs.map((i) => `### ${i.reference}\n${safeJson(i.value)}`).join("\n\n")}`
+        : "";
+    // Gate-loop feedback (R2 max_loops): the judge's rejection is appended so
+    // the re-executed unit can address it — and so the input hash changes,
+    // making the loop's re-dispatch natural instead of a durable-row reuse.
+    const gateBlock = gateFeedback
+        ? `\n\n## Completion-gate feedback (previous attempt rejected)\n` +
+            `A completion-criteria judge rejected this step's previous results. Address this feedback:\n` +
+            gateFeedback.feedback +
+            (gateFeedback.missing.length > 0
+                ? `\nUnmet criteria:\n${gateFeedback.missing.map((m) => `- ${m}`).join("\n")}`
+                : "")
+        : "";
+    const schemaDirective = schema
+        ? `\n\nRespond with ONLY a JSON value matching this JSON Schema (no prose, no code fences):\n${safeJson(schema)}`
+        : "";
+    return `${preamble}\n${instructions}${itemBlock}${inputsBlock}${gateBlock}${schemaDirective}`;
+}
+/**
+ * Content-derived unit identity (module doc): `<node_id>:<hash12>` for a
+ * fan-out item, `<node_id>:solo` otherwise. The hash is over the item's
+ * canonical JSON (sorted keys — same canonicalization the vote reducer
+ * counts with), so identity survives list reordering/regeneration and is
+ * independent of item position. Retry attempts stack `~r<n>` on top.
+ */
+export function unitIdFor(nodeId, item, isFanOut) {
+    if (!isFanOut)
+        return `${nodeId}:solo`;
+    const canonical = canonicalJson(item) ?? "null";
+    return `${nodeId}:${createHash("sha256").update(canonical).digest("hex").slice(0, 12)}`;
+}
+/** Include an SDK fallback in v3 call identity without copying catalog entries onto nodes. */
+function transitiveDispatchSnapshot(engine, engines) {
+    if (engine.kind !== "agent" || !engine.fallbackLlmEngine)
+        return engine;
+    const fallback = engines[engine.fallbackLlmEngine];
+    if (!fallback || fallback.kind !== "llm") {
+        throw new UsageError(`Frozen agent engine "${engine.name}" has no valid LLM fallback snapshot.`);
+    }
+    return { engine, fallback };
+}
+// ── Step outputs + reducers + typed artifacts ────────────────────────────────
+/**
+ * The value `${{ steps.<id>.output }}` resolves to for ONE step, given that
+ * step's journaled evidence: an engine-executed step carries a promoted
+ * ARTIFACT under `evidence.output` (solo unit result/text, collect array, or
+ * vote winner); evidence without an `output` key (manually-completed steps) is
+ * exposed as-is.
+ */
+export function projectStepOutput(evidence) {
+    return Object.hasOwn(evidence, "output") ? evidence.output : evidence;
+}
+/** Project the engine's evidence map into the expression scope's `stepOutputs`. */
+export function stepOutputsFromEvidence(evidence) {
+    const outputs = {};
+    for (const [stepId, stepEvidence] of Object.entries(evidence)) {
+        if (stepEvidence !== undefined)
+            outputs[stepId] = projectStepOutput(stepEvidence);
+    }
+    return outputs;
+}
+/**
+ * Typed artifacts (addendum, R2): validate the promoted step artifact against
+ * `IrStepPlan.outputSchema`. Returns the step-failure summary (validation
+ * errors included) on mismatch, undefined when valid or when no schema is
+ * declared.
+ */
+export function validateStepArtifact(plan, evidence) {
+    if (!plan.outputSchema)
+        return undefined;
+    const errors = validateJsonSchemaSubset(projectStepOutput(evidence), plan.outputSchema);
+    if (errors.length === 0)
+        return undefined;
+    return (`Step "${plan.stepId}" artifact failed validation against the step's declared output schema: ` +
+        `${errors.join("; ")}.`);
+}
+/**
+ * Build the summary the completion-criteria gate judges for a step (addendum
+ * R2, "typed artifacts, honest gates"): a one-line unit count followed by the
+ * promoted step artifact as canonical JSON, clipped at {@link GATE_ARTIFACT_CLIP}
+ * chars. This replaces machine-prose so the gate evaluates real results.
+ */
+export function buildArtifactSummary(stepId, units, evidence) {
+    const failedCount = units.filter((u) => !u.ok).length;
+    const json = canonicalJson(projectStepOutput(evidence)) ?? "null";
+    return (`Step "${stepId}" executed ${units.length} unit(s) (${units.length - failedCount} succeeded, ${failedCount} failed). ` +
+        `Step artifact (canonical JSON${json.length > GATE_ARTIFACT_CLIP ? `, clipped at ${GATE_ARTIFACT_CLIP} chars` : ""}):\n` +
+        clip(json, GATE_ARTIFACT_CLIP));
+}
+/** A unit's contribution to the step artifact: structured result, else text, else null (failures). */
+function unitOutputValue(unit) {
+    if (!unit.ok)
+        return null;
+    if (unit.result !== undefined)
+        return unit.result;
+    return unit.text ?? null;
+}
+export function buildEvidence(units, reducer, isFanOut) {
+    // Per-unit evidence is the DURABLE, surface-independent projection the two
+    // driver surfaces (engine + brief/report) must agree on byte-for-byte (R4
+    // conformance, "identical unit graph"). It therefore carries ONLY fields both
+    // surfaces can reproduce from the journal:
+    //   - a SUCCESS keeps its promoted contribution (structured `result` or clipped
+    //     `text`) — the report path rehydrates exactly these from the unit row;
+    //   - a FAILURE keeps only its `failureReason` (the durable, journaled failure
+    //     vocabulary). The engine's in-memory dispatch diagnostic (`error`) and any
+    //     residual `text` on a failed unit are NOT persisted here: a driver-reported
+    //     failure carries neither, so persisting them on the engine surface alone
+    //     would diverge the durable graph. The full raw text/reason still lives on
+    //     the unit row for engine-side diagnostics; this is the shared graph.
+    const collected = units.map((u) => u.ok
+        ? {
+            unitId: u.unitId,
+            ok: true,
+            ...(u.result !== undefined ? { result: u.result } : {}),
+            ...(u.text !== undefined ? { text: clip(u.text, EVIDENCE_TEXT_CLIP) } : {}),
+        }
+        : {
+            unitId: u.unitId,
+            ok: false,
+            ...(u.failureReason ? { failureReason: u.failureReason } : {}),
+        });
+    const evidence = { units: collected, itemCount: units.length };
+    // Promoted step artifact (`evidence.output`) — what `${{ steps.<id>.output }}`
+    // resolves to (see projectStepOutput). Values are UNCLIPPED.
+    if (reducer === "vote") {
+        evidence.output = null;
+    }
+    else {
+        evidence.output = isFanOut ? units.map(unitOutputValue) : unitOutputValue(units[0]);
+    }
+    if (reducer === "vote") {
+        const counts = new Map();
+        for (const unit of units) {
+            if (!unit.ok)
+                continue;
+            const value = unit.result !== undefined ? unit.result : unit.text;
+            const key = canonicalJson(value);
+            const entry = counts.get(key);
+            if (entry)
+                entry.count++;
+            else
+                counts.set(key, { value, count: 1 });
+        }
+        const ranked = [...counts.values()].sort((a, b) => b.count - a.count);
+        if (ranked.length === 0) {
+            evidence.voteError = "Vote reducer had no successful unit results to count.";
+        }
+        else if (ranked.length > 1 && ranked[0].count === ranked[1].count) {
+            evidence.voteError = `Vote reducer tied at ${ranked[0].count} vote(s) — no majority.`;
+        }
+        else {
+            evidence.vote = { winner: ranked[0].value, votes: ranked[0].count, total: units.length };
+            evidence.output = ranked[0].value;
+        }
+    }
+    return evidence;
+}
+/**
+ * Reduce a step's terminal unit outcomes into the promoted artifact + step
+ * verdict — the shared semantics between native dispatch and the report path.
+ * Applies the `on_error` policy (`fail` vs `continue`), the reducer (via
+ * {@link buildEvidence}), the vote-tie failure, and the typed-artifact schema
+ * validation (fail-fast, errors in the summary, `artifactSchemaFailure` marker).
+ * Callers own dispatch-specific concerns (replay-divergence, budget) BEFORE
+ * calling this; those never occur on the report path (units are journaled).
+ */
+export function reduceStepOutcomes(plan, reducer, isFanOut, onError, units) {
+    const failed = units.filter((u) => !u.ok);
+    const evidence = buildEvidence(units, reducer, isFanOut);
+    const reducerNote = typeof evidence.voteError === "string" ? ` ${evidence.voteError}` : "";
+    const tolerateFailures = onError === "continue";
+    let ok = (tolerateFailures || failed.length === 0) && !evidence.voteError;
+    let summary = `Executed ${units.length} unit(s) for step "${plan.stepId}" via workflow orchestration: ` +
+        `${units.length - failed.length} succeeded, ${failed.length} failed.` +
+        (failed.length > 0
+            ? ` Failures${tolerateFailures ? " (recorded, on_error: continue)" : ""}: ${failed
+                .map((u) => `${u.unitId} (${u.failureReason ?? "error"})`)
+                .join(", ")}.`
+            : "") +
+        reducerNote;
+    let artifactSchemaFailure = false;
+    if (ok) {
+        const schemaFailure = validateStepArtifact(plan, evidence);
+        if (schemaFailure !== undefined) {
+            ok = false;
+            summary = schemaFailure;
+            artifactSchemaFailure = true;
+        }
+    }
+    return { ok, units, evidence, summary, ...(artifactSchemaFailure ? { artifactSchemaFailure: true } : {}) };
+}
+/**
+ * The reduced outcome of a step whose fan-out list resolved to EMPTY (`over: []`
+ * or a producer that yielded `[]`): no units are dispatched, so the promoted
+ * artifact is the degenerate empty value — the empty array for a `collect`
+ * reducer, `null` for `vote` (references into a missing winner fail loudly at
+ * resolution rather than silently reading the envelope). Even the degenerate
+ * artifact must honor the step's declared `outputSchema` before it can complete.
+ *
+ * Shared by native dispatch (`executeStepPlan`'s `items.length === 0` branch)
+ * and the R3 driver protocol (`report` auto-completes an empty step the spine
+ * reaches, since no `report --unit` can ever advance a zero-unit step) so both
+ * surfaces promote the SAME artifact and apply the SAME schema verdict — the
+ * anti-drift guarantee. Deliberately does NOT run the reducer/vote-tie logic:
+ * an empty step has no successful results to count, and a vote-tie "failure"
+ * would diverge from the engine's long-standing empty-list semantics.
+ */
+export function reduceEmptyStep(plan, reducer) {
+    const evidence = { units: [], itemCount: 0, output: reducer === "collect" ? [] : null };
+    const schemaFailure = validateStepArtifact(plan, evidence);
+    return {
+        ok: schemaFailure === undefined,
+        units: [],
+        evidence,
+        summary: schemaFailure ?? `Step "${plan.stepId}" fan-out list was empty — no units dispatched.`,
+        ...(schemaFailure !== undefined ? { artifactSchemaFailure: true } : {}),
+    };
+}
+/**
+ * Rehydrate a journaled unit row into a {@link UnitOutcome}. Shared by the
+ * executor's durable-row reuse (`native-executor.ts`, completed rows only) and
+ * the R3 report path (which reduces completed AND failed rows replayed from the
+ * journal). A completed row's text unit journals its output as a JSON string; a
+ * schema unit journals the validated structure. A failed row carries its
+ * `failure_reason`; any journaled text is surfaced too.
+ */
+export function unitOutcomeFromRow(unitId, row, hasSchema) {
+    let parsed;
+    try {
+        parsed = row.result_json === null ? undefined : JSON.parse(row.result_json);
+    }
+    catch {
+        parsed = undefined;
+    }
+    if (row.status === "completed") {
+        return {
+            unitId,
+            ok: true,
+            ...(hasSchema
+                ? { result: parsed }
+                : typeof parsed === "string"
+                    ? { text: parsed }
+                    : parsed !== undefined
+                        ? { result: parsed }
+                        : {}),
+            ...(row.tokens !== null ? { tokens: row.tokens } : {}),
+            ...(row.session_id !== null && row.session_id !== undefined ? { sessionId: row.session_id } : {}),
+        };
+    }
+    return {
+        unitId,
+        ok: false,
+        failureReason: row.failure_reason ?? "reported_failure",
+        ...(typeof parsed === "string" ? { text: parsed } : {}),
+        ...(row.tokens !== null ? { tokens: row.tokens } : {}),
+    };
+}
+/**
+ * Select the journaled attempt row that determines a unit's TERMINAL outcome on
+ * a REPLAY surface — the engine's durable-row reuse AND the harness-neutral
+ * brief/report driver protocol — given the run's dispatch rows indexed by
+ * unit_id. This is the ONE place all surfaces resolve "which journaled row IS
+ * this unit's outcome," so they cannot drift from each other or from the engine.
+ *
+ * It mirrors the executor's {@link classifyUnitReuse} attempt scan
+ * (native-executor.ts): among the base attempt and its `~r<n>` retries — all
+ * stacked on `journalBaseId`, which already carries the active `~l<loop>` gate
+ * suffix — the FIRST completed attempt is the effective result. So a unit whose
+ * base attempt FAILED but whose later retry COMPLETED reduces as COMPLETED,
+ * exactly like an engine resume reusing the `~r1` row (Codex round-3 finding C);
+ * reading only the base row would reduce it as failed and diverge the two
+ * surfaces. With no completed attempt the HIGHEST journaled attempt stands (a
+ * terminal failure, or a still-running row); no attempt row at all ⇒ `undefined`
+ * (the unit is still outstanding).
+ */
+export function selectUnitAttemptRow(workUnit, dispatchRows) {
+    const base = workUnit.journalBaseId;
+    const maxAttempts = 1 + Math.max(0, workUnit.retry?.max ?? 0);
+    let fallback;
+    for (let attempt = 0; attempt < maxAttempts; attempt++) {
+        const row = dispatchRows.get(attempt === 0 ? base : `${base}~r${attempt}`);
+        if (!row)
+            continue;
+        if (row.status === "completed")
+            return row;
+        fallback = row; // remember the highest journaled (non-completed) attempt
+    }
+    return fallback;
+}
+/**
+ * Is a FAILED unit still RETRY-ELIGIBLE — i.e. NOT terminal, because a driver
+ * could still re-run it via the `--rerun` form (the engine's automatic
+ * `<baseId>~r<n>` retry)? A unit whose declared `retry.on` matches the recorded
+ * failure reason AND whose attempt budget (`1 + retry.max`) is not yet spent can
+ * still be re-run. No `retry`, an off-list reason, or an exhausted attempt budget
+ * ⇒ the failure IS terminal. Shared by the report fail-fast decision, the
+ * `--settle` refusal, and `brief`'s fully-terminal detection so all three agree
+ * on when a failed unit is genuinely done vs. still re-runnable. The normalized
+ * failure reason is compared against `retry.on` directly (a canonical taxonomy
+ * reason is stored verbatim; an `external:*` reason is by construction outside
+ * the taxonomy `retry.on` lists).
+ */
+export function isRetryEligibleFailure(workUnit, row, failureReason) {
+    const retry = workUnit.retry;
+    if (!retry || failureReason === null || !retry.on.includes(failureReason))
+        return false;
+    const attempts = row?.attempts ?? 1;
+    return attempts < 1 + Math.max(0, retry.max);
+}
+/**
+ * Does a resolvable unit still need a driver to execute + report it (or re-run
+ * it)? True for a unit with no terminal row (pending), a still-`running` row (a
+ * live/stale claim another driver holds), or a FAILED row that is still
+ * retry-eligible. False for a COMPLETED row, a terminal non-retry-eligible
+ * FAILURE, or an UNRESOLVABLE unit (a currently-unreachable defensive branch —
+ * see the computeStepWorkList doc — never reportable). The best terminal attempt (base + `~r<n>` retries) is the
+ * one consulted, the SAME reuse the engine and reducer apply.
+ */
+export function unitStillNeedsReport(workUnit, dispatchRows) {
+    if (!workUnit.resolved.ok)
+        return false;
+    const row = selectUnitAttemptRow(workUnit, dispatchRows);
+    if (!row)
+        return true; // no journal row → pending
+    if (row.status === "running")
+        return true; // a live/stale claim is still in flight
+    if (row.status === "failed")
+        return isRetryEligibleFailure(workUnit, row, row.failure_reason);
+    return false; // completed (or a non-retry-eligible failure) → terminal
+}
+/**
+ * Is the active step's work-list FULLY TERMINAL — every resolvable unit run to a
+ * terminal (done, or non-retry-eligible failed) state with nothing left to
+ * execute or per-unit report — yet still needing finalization? This is the
+ * driver-recovery state after a required-gate block is resumed, or a crash
+ * between the last unit write and the step's completion (owner manual-validation
+ * finding 3): the work-list is done but the step never advanced. `brief`
+ * surfaces it with a single `report --settle` command and `--settle` runs the
+ * shared completion path for it. A list with ANY outstanding unit (pending,
+ * in-flight, or retry-eligible failed) is NOT fully terminal — the driver
+ * `report --unit`s those. A route-only / empty / all-unresolvable list (no
+ * resolvable units) is a DIFFERENT non-dispatching state, handled separately.
+ */
+export function isWorkListFullyTerminal(workList, dispatchRows) {
+    if (!workList.units.some((u) => u.resolved.ok))
+        return false;
+    return workList.units.every((u) => !unitStillNeedsReport(u, dispatchRows));
+}
+/** Stable stringify (sorted object keys, recursively) so equal values vote together. */
+export function canonicalJson(value) {
+    return JSON.stringify(sortKeys(value));
+}
+function sortKeys(value) {
+    if (Array.isArray(value))
+        return value.map(sortKeys);
+    if (value && typeof value === "object") {
+        return Object.fromEntries(Object.entries(value)
+            .sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))
+            .map(([k, v]) => [k, sortKeys(v)]));
+    }
+    return value;
+}
+// ── Gate-feedback recovery (PURE) ────────────────────────────────────────────
+//
+// A gate rejection is journaled as `<stepId>.gate:l<loop>` with result_json
+// `{ complete: false, missing, feedback }` (see journalGateEvaluationFinish).
+// The feedback stored there is BYTE-IDENTICAL to what the engine threads into
+// the next loop's prompts — both are the same `rejection.feedback`/`.missing`.
+// `brief` recovers it from the journal so its loop-N work-list matches the
+// engine's (redesign addendum R3, task item 2). `native-executor.test.ts`
+// asserts the round-trip identity.
+// GATE_EVALUATION_PHASE moved to ../runtime/unit-phases.ts (leaf) so
+// unit-checkin can key on it without closing the exec ↔ runtime cycle.
+/** The unit id of a step's gate-evaluation row for a given 1-based loop. */
+export function gateUnitId(stepId, loop) {
+    return `${stepId}.gate:l${loop}`;
+}
+/**
+ * The gate loop the engine is about to (re-)run for an ACTIVE step, derived
+ * purely from the journal: one past the highest journaled loop that REJECTED
+ * (`complete: false`). No rejected gate rows ⇒ loop 1 (the first execution).
+ * A passed gate would have advanced the spine, so an active step never has a
+ * `complete: true` row as its latest gate evaluation.
+ *
+ * Reviewer #17: a gate row that EXISTS but cannot be parsed (or carries an
+ * invalid verdict shape) is CORRUPTION — {@link parseGateVerdict} throws loudly
+ * rather than letting `gateRowRejected` swallow the parse error, which would
+ * silently drop the loop back to 1 and re-dispatch work whose gate outcome is
+ * unknown.
+ */
+export function activeGateLoop(rows, stepId) {
+    let maxRejectedLoop = 0;
+    for (const row of rows) {
+        if (row.phase !== GATE_EVALUATION_PHASE || row.step_id !== stepId)
+            continue;
+        const loop = gateLoopOf(row.unit_id, stepId);
+        if (loop === undefined)
+            continue;
+        // Throws loudly on a corrupt/malformed gate row — never treated as absent.
+        if (parseGateVerdict(row).kind === "rejected" && loop > maxRejectedLoop)
+            maxRejectedLoop = loop;
+    }
+    return maxRejectedLoop + 1;
+}
+/**
+ * Recover the gate feedback the engine threads into `loop`'s unit prompts: the
+ * `{ feedback, missing }` journaled by the previous loop's rejection
+ * (`<stepId>.gate:l<loop-1>`). Loop 1 (or a missing/passed/errored previous row)
+ * has no feedback. Pure — the journal rows are passed in.
+ *
+ * Reviewer #17: a PRESENT previous gate row that cannot be parsed fails LOUDLY
+ * (via {@link parseGateVerdict}) instead of returning undefined — a corrupt row
+ * must not make an in-loop step look like loop 1 with no recovered feedback.
+ */
+export function recoverGateFeedback(rows, stepId, loop) {
+    if (loop <= 1)
+        return undefined;
+    const prevId = gateUnitId(stepId, loop - 1);
+    const prev = rows.find((r) => r.unit_id === prevId && r.phase === GATE_EVALUATION_PHASE);
+    if (!prev)
+        return undefined;
+    const verdict = parseGateVerdict(prev);
+    return verdict.kind === "rejected" ? { feedback: verdict.feedback, missing: verdict.missing } : undefined;
+}
+/** The 1-based loop encoded in a `<stepId>.gate:l<n>` unit id, if well-formed. */
+function gateLoopOf(unitId, stepId) {
+    const prefix = `${stepId}.gate:l`;
+    if (!unitId.startsWith(prefix))
+        return undefined;
+    const n = Number.parseInt(unitId.slice(prefix.length), 10);
+    return Number.isInteger(n) && n >= 1 ? n : undefined;
+}
+/**
+ * Classify a gate-evaluation row's journaled verdict, failing LOUDLY on a
+ * corrupt one (reviewer #17). A NULL `result_json` is the LEGITIMATE
+ * errored-judge / in-flight shape (`journalGateEvaluationFinish` writes null for
+ * an errored judge, and a `running` row has no verdict yet) and classifies as
+ * `empty`. But a PRESENT `result_json` that does not parse as JSON, or parses to
+ * anything other than an object with a boolean `complete` field, is corruption —
+ * a truncated or hand-edited row — and MUST NOT be silently treated as absent
+ * (which would reset an active step's gate loop to 1 and re-dispatch work whose
+ * completion outcome is unknown). We refuse to guess.
+ */
+function parseGateVerdict(row) {
+    if (row.result_json === null)
+        return { kind: "empty" };
+    let verdict;
+    try {
+        verdict = JSON.parse(row.result_json);
+    }
+    catch {
+        throw new UsageError(gateCorruptionMessage(row, "its result_json is not valid JSON"));
+    }
+    if (typeof verdict !== "object" || verdict === null || Array.isArray(verdict)) {
+        throw new UsageError(gateCorruptionMessage(row, "its result_json is not a JSON object"));
+    }
+    const v = verdict;
+    if (typeof v.complete !== "boolean") {
+        throw new UsageError(gateCorruptionMessage(row, 'its verdict has no boolean "complete" field'));
+    }
+    if (v.complete === false) {
+        const feedback = typeof v.feedback === "string" ? v.feedback : "";
+        const missing = Array.isArray(v.missing) ? v.missing.filter((m) => typeof m === "string") : [];
+        return { kind: "rejected", missing, feedback };
+    }
+    return { kind: "passed" };
+}
+function gateCorruptionMessage(row, why) {
+    return (`Workflow run ${row.run_id} has a corrupt gate-evaluation row "${row.unit_id}" for step "${row.step_id}" — ${why}. ` +
+        `A gate verdict must be {"complete": true|false, …}; refusing to treat a malformed gate row as absent, which would ` +
+        `silently restart the step's gate loop and re-dispatch work whose completion outcome is unknown. Fix or remove the ` +
+        `journaled row, then resume the run.`);
+}
+/** Insert the gate-evaluation unit row (running) just before the judge runs. */
+export async function journalGateEvaluationStart(gate) {
+    const unitId = gateUnitId(gate.stepId, gate.loop);
+    await enqueueUnitWrite(() => withWorkflowRunsRepo((repo) => repo.insertUnit({
+        runId: gate.runId,
+        unitId,
+        stepId: gate.stepId,
+        nodeId: `${gate.stepId}.gate`,
+        parentUnitId: null,
+        // Marks the row as a judge call, NOT a dispatch: the budget/lifetime
+        // seed in `driveRun` skips these so resume accounting matches live.
+        phase: GATE_EVALUATION_PHASE,
+        runner: "llm",
+        engine: gate.invocation.engine,
+        model: gate.invocation.model,
+        inputHash: gate.inputHash,
+        startedAt: new Date().toISOString(),
+    })));
+    appendEvent({
+        eventType: "workflow_unit_started",
+        ref: gate.workflowRef,
+        metadata: { runId: gate.runId, stepId: gate.stepId, unitId },
+    });
+}
+/**
+ * Finish the gate-evaluation unit row with the verdict as observed from the
+ * completion outcome: a rejection journals `{ complete: false, missing,
+ * feedback }`; a pass journals `{ complete: true, missing: [] }`; a judge that
+ * threw journals a failed row with a NULL verdict and validation is skipped.
+ */
+export async function journalGateEvaluationFinish(gate, errored, rejection) {
+    const unitId = gateUnitId(gate.stepId, gate.loop);
+    const verdict = errored
+        ? null
+        : rejection
+            ? { complete: false, missing: rejection.missing, feedback: rejection.feedback }
+            : { complete: true, missing: [] };
+    const status = errored ? "failed" : "completed";
+    await enqueueUnitWrite(() => withWorkflowRunsRepo((repo) => repo.finishUnit({
+        runId: gate.runId,
+        unitId,
+        status,
+        resultJson: verdict ? JSON.stringify(verdict) : null,
+        tokens: null,
+        failureReason: errored ? "dispatch_error" : null,
+        finishedAt: new Date().toISOString(),
+    })));
+    appendEvent({
+        eventType: "workflow_unit_finished",
+        ref: gate.workflowRef,
+        metadata: { runId: gate.runId, stepId: gate.stepId, unitId, status },
+    });
+}
+/**
+ * Resolve a route's input (a single whole-value `${{ … }}` reference) and pick
+ * the branch. No ambient key search. Only primitive values route; the
+ * comparison is exact string equality against the declared `when:` matches.
+ */
+export function evaluateRoute(route, scope) {
+    const resolved = resolveReferenceString(route.input, scope);
+    if (!resolved.ok) {
+        return { ok: false, error: `route input ${route.input} failed to resolve: ${resolved.error.message}` };
+    }
+    const value = resolved.value;
+    if (typeof value === "object" && value !== null) {
+        return {
+            ok: false,
+            error: `route input ${route.input} resolved to a non-primitive value; branches match on strings/numbers/booleans.`,
+        };
+    }
+    const valueString = typeof value === "string" ? value : String(value);
+    // Own-property check: `when` is author-controlled, and a value such as
+    // "constructor" must not resolve through Object.prototype.
+    const selected = Object.hasOwn(route.when, valueString) ? route.when[valueString] : route.defaultStepId;
+    if (!selected) {
+        return {
+            ok: false,
+            error: `value "${valueString}" matched no "when:" branch and the route declares no default.`,
+        };
+    }
+    return { ok: true, value: valueString, selected };
+}
+/**
+ * Cascade a SKIPPED router: it never evaluated its route, so every declared
+ * target (branches + default) is marked skip-on-reach unless an earlier router
+ * already claimed it. Shared by the live skip path and the journal replay.
+ */
+export function cascadeSkippedRouter(route, routerId, routeUnselected) {
+    const targets = [...Object.values(route.when), ...(route.defaultStepId ? [route.defaultStepId] : [])];
+    for (const target of targets) {
+        if (!routeUnselected.has(target)) {
+            routeUnselected.set(target, { router: routerId, selected: null });
+        }
+    }
+}
+/**
+ * Record one router's decision in the skip bookkeeping: the selected target is
+ * protected, every other declared target (branches + default) is marked
+ * skip-on-reach unless an earlier router already claimed it. Shared by the live
+ * evaluation path and the journal replay.
+ */
+export function applyRouteDecision(route, routerId, selected, routeSelected, routeUnselected) {
+    routeSelected.add(selected);
+    const targets = [...Object.values(route.when), ...(route.defaultStepId ? [route.defaultStepId] : [])];
+    for (const target of targets) {
+        if (target !== selected && !routeUnselected.has(target)) {
+            routeUnselected.set(target, { router: routerId, selected });
+        }
+    }
+}
+/**
+ * The `stepOutputs` scope a route resolves against: every prior step's recorded
+ * evidence plus the just-finished step's fresh evidence — each projected
+ * through {@link projectStepOutput}. Same projection as unit templates, so the
+ * two scopes cannot drift.
+ */
+export function routeStepOutputs(evidence, currentStepId, currentEvidence) {
+    const outputs = {};
+    for (const [stepId, stepEvidence] of Object.entries(evidence)) {
+        if (stepEvidence !== undefined)
+            outputs[stepId] = projectStepOutput(stepEvidence);
+    }
+    outputs[currentStepId] = projectStepOutput(currentEvidence);
+    return outputs;
+}
+/** The `selected` target journaled on a route step's evidence, if well-formed. */
+function journaledRouteSelection(evidence) {
+    const route = evidence?.route;
+    if (typeof route !== "object" || route === null || Array.isArray(route))
+        return undefined;
+    const selected = route.selected;
+    return typeof selected === "string" && selected !== "" ? selected : undefined;
+}
+/** The set of steps a route may legally select: its `when` branches + default. */
+function routeTargets(route) {
+    return new Set([...Object.values(route.when), ...(route.defaultStepId ? [route.defaultStepId] : [])]);
+}
+/**
+ * Reviewer #7: a journaled route decision must name a target the route actually
+ * DECLARES (`when` branch or `default`). Corrupted or hand-edited evidence can
+ * otherwise mark a non-existent step as `selected` — which unselects and skips
+ * every REAL branch target, silently steering the run down a phantom branch.
+ * `evaluateRoute` can only ever produce a declared target, so a stored value
+ * outside that set is provably tampered evidence: fail loudly rather than seed a
+ * bogus skip set.
+ */
+function assertRouteTargetDeclared(route, stepId, selected, runId) {
+    const targets = routeTargets(route);
+    if (!targets.has(selected)) {
+        throw new UsageError(`Workflow run ${runId} has a completed route step "${stepId}" whose journaled route decision selected ` +
+            `"${selected}", which is not a declared branch or default target of the route (valid targets: ` +
+            `${[...targets].join(", ") || "(none)"}). The route evidence was corrupted or manually edited — refusing to ` +
+            `apply a bogus route decision that would skip the real branch targets. Start a new run.`);
+    }
+}
+/**
+ * Validate every COMPLETED route step's journaled selection against its declared
+ * targets (reviewer #7). Read-only: it throws on a PRESENT-but-invalid selection
+ * and is silent on an absent one, so it never false-positives on a healthy run —
+ * making it safe to call from the read-only `brief` surface as well as the
+ * resume/report surfaces that already re-apply the decisions.
+ */
+export function assertJournaledRouteSelectionsValid(plan, state) {
+    for (const stepPlan of plan.steps) {
+        if (!stepPlan.route)
+            continue;
+        const stepState = state.workflow.steps.find((s) => s.id === stepPlan.stepId);
+        if (!stepState || stepState.status !== "completed")
+            continue;
+        const selected = journaledRouteSelection(stepState.evidence);
+        if (selected !== undefined) {
+            assertRouteTargetDeclared(stepPlan.route, stepPlan.stepId, selected, state.run.id);
+        }
+    }
+}
+/**
+ * Replay journaled route decisions into the skip bookkeeping (resume path).
+ * For every COMPLETED route step of the frozen plan, in spine order: the
+ * journaled decision wins; else a re-derivation from the frozen plan +
+ * journaled evidence; else fail loudly. A SKIPPED route step cascades its
+ * targets into the skip set exactly as on the live path.
+ */
+export function seedJournaledRouteDecisions(plan, state, routeSelected, routeUnselected) {
+    const evidence = {};
+    for (const s of state.workflow.steps)
+        evidence[s.id] = s.evidence;
+    for (const stepPlan of plan.steps) {
+        if (!stepPlan.route)
+            continue;
+        const stepState = state.workflow.steps.find((s) => s.id === stepPlan.stepId);
+        if (!stepState)
+            continue;
+        if (stepState.status === "skipped") {
+            cascadeSkippedRouter(stepPlan.route, stepPlan.stepId, routeUnselected);
+            continue;
+        }
+        if (stepState.status !== "completed")
+            continue;
+        let selected = journaledRouteSelection(stepState.evidence);
+        if (selected !== undefined) {
+            // Reviewer #7: a stored decision must name a declared target — a bogus one
+            // (tampered/hand-edited evidence) fails loudly rather than seeding a skip
+            // set that buries the real branches.
+            assertRouteTargetDeclared(stepPlan.route, stepPlan.stepId, selected, state.run.id);
+        }
+        if (selected === undefined) {
+            const scope = {
+                params: state.run.params ?? {},
+                stepOutputs: routeStepOutputs(evidence, stepPlan.stepId, stepState.evidence ?? {}),
+            };
+            const decision = evaluateRoute(stepPlan.route, scope);
+            if (decision.ok)
+                selected = decision.selected;
+        }
+        if (selected === undefined) {
+            throw new UsageError(`Workflow run ${state.run.id} has a completed route step "${stepPlan.stepId}" with no journaled route ` +
+                `decision, and the decision cannot be re-derived from the journaled evidence. Refusing to guess which ` +
+                `branch was selected — advance the remaining steps manually with \`akm workflow complete\`.`);
+        }
+        applyRouteDecision(stepPlan.route, stepPlan.stepId, selected, routeSelected, routeUnselected);
+    }
+}
+/**
+ * Perform ONE completion attempt for an executed step:
+ *
+ *  - a hard unit failure completes the step `failed` (a retryable typed-artifact
+ *    mismatch with loops remaining returns `retry` WITHOUT journaling a gate row
+ *    — no judge ran, exactly like the engine);
+ *  - a route decision is evaluated against params + prior/fresh step outputs; an
+ *    unroutable value fails the step; a valid decision is journaled on the
+ *    step evidence and applied to the skip bookkeeping;
+ *  - the completion gate judges a summary BUILT FROM the promoted artifact (when
+ *    the step declares criteria), journaled as a `<stepId>.gate:l<loop>` unit
+ *    row; a rejection with loops remaining returns `retry` (feedback threaded
+ *    into the next loop), a rejection with none returns `gate-exhausted`, a pass
+ *    returns `advanced`.
+ *
+ * Every DB advance goes through {@link completeWorkflowStep} — the gate spine is
+ * never bypassed. Behavior is byte-identical to the engine's former inline loop
+ * body (its tests prove it).
+ */
+export async function finalizeExecutedStep(input) {
+    const { runId, workflowRef, stepId, stepPlan, completionCriteria, gateLoop, loopsRemaining, result } = input;
+    const lease = input.leaseHolder !== undefined ? { leaseHolder: input.leaseHolder } : {};
+    if (!result.ok) {
+        // Typed-artifact mismatch with loop budget left: regenerate-with-errors
+        // (the validation errors become the next loop's feedback). No judge ran, so
+        // no gate row is journaled for this attempt.
+        if (result.artifactSchemaFailure && loopsRemaining) {
+            return { kind: "retry", gateFeedback: { feedback: result.summary, missing: [] } };
+        }
+        await completeWorkflowStep({
+            runId,
+            stepId,
+            status: "failed",
+            notes: result.summary,
+            evidence: result.evidence,
+            ...lease,
+        });
+        return { kind: "failed", summary: result.summary };
+    }
+    // Resolve the completion-criteria judge ONCE (reused by the gate below). A
+    // A frozen plan either supplies its judge at the dispatch boundary or has no
+    // judge. Re-selecting defaults here would let config drift change a run.
+    const innerJudge = input.summaryJudge ?? null;
+    // Route evaluation BEFORE completion: an unroutable value is an
+    // authoring/config failure that must fail the step deterministically.
+    let summaryOverride;
+    if (stepPlan.route) {
+        const scope = {
+            params: input.params,
+            stepOutputs: routeStepOutputs(input.priorEvidence, stepId, result.evidence),
+        };
+        const decision = evaluateRoute(stepPlan.route, scope);
+        if (!decision.ok) {
+            const notes = `Step "${stepId}" route failed: ${decision.error}`;
+            await completeWorkflowStep({ runId, stepId, status: "failed", notes, evidence: result.evidence, ...lease });
+            return { kind: "failed", summary: notes, routeFailure: true };
+        }
+        applyRouteDecision(stepPlan.route, stepId, decision.selected, input.routeSelected, input.routeUnselected);
+        // Journal the decision on the evidence: resume replays it via
+        // seedJournaledRouteDecisions, so the skip set survives re-invocation.
+        result.evidence.route = { input: stepPlan.route.input, value: decision.value, selected: decision.selected };
+        if (!stepPlan.root) {
+            summaryOverride = `Step "${stepId}" routed on ${stepPlan.route.input}: value "${decision.value}" selected step "${decision.selected}".`;
+        }
+    }
+    // Artifact-judging gate: a criteria-bearing executing step is judged on a
+    // summary BUILT FROM the promoted artifact; everything else keeps the machine
+    // summary (a route-only step's summary IS its decision).
+    const summary = stepPlan.root && completionCriteria.length > 0
+        ? buildArtifactSummary(stepId, result.units, result.evidence)
+        : (summaryOverride ?? result.summary);
+    // Journal engine-driven judge calls as unit rows (they are LLM calls). The
+    // wrapper's `invoked` stays false when the gate is fail-open (no criteria / no
+    // judge) — nothing is journaled, and human approvals are never cached.
+    const frozenGate = innerJudge
+        ? await withWorkflowRunsRepo((repo) => {
+            const row = repo.getRunById(runId);
+            if (!row)
+                throw new UsageError(`Workflow run ${runId} was not found.`);
+            const plan = requireExecutableWorkflowPlan(row);
+            const invocation = plan.steps.find((step) => step.stepId === stepId)?.gate.judge ?? null;
+            return invocation ? { invocation, engine: plan.execution?.engines[invocation.engine] ?? null } : null;
+        })
+        : null;
+    const gateInvocation = frozenGate?.invocation ?? null;
+    let gateUnit;
+    const judgeState = { invoked: false, errored: false };
+    const summaryJudge = innerJudge
+        ? async (prompt) => {
+            judgeState.invoked = true;
+            if (gateInvocation) {
+                gateUnit = {
+                    runId,
+                    workflowRef,
+                    stepId,
+                    loop: gateLoop,
+                    invocation: gateInvocation,
+                    inputHash: createHash("sha256")
+                        .update(canonicalJsonString({
+                        hashVersion: 3,
+                        dispatch: frozenGate?.engine ?? null,
+                        invocation: gateInvocation,
+                        prompt,
+                    }))
+                        .digest("hex"),
+                };
+                await journalGateEvaluationStart(gateUnit);
+            }
+            try {
+                return await innerJudge(prompt);
+            }
+            catch (err) {
+                judgeState.errored = true;
+                throw err;
+            }
+        }
+        : null;
+    // Reviewer #6: once the judge is invoked, its gate row is journaled `running`
+    // (journalGateEvaluationStart) and MUST be finished on every exit. The
+    // already-fixed window is the judge itself throwing (caught inside
+    // validateStepSummary — `judgeState.errored` records it). The remaining
+    // window is `completeWorkflowStep` throwing AFTER the judge ran — a stolen
+    // lease, a concurrent state change, a DB error — which would otherwise skip the
+    // finish and strand the gate row in `running`. Finish it as an errored row (the
+    // observed outcome: the completion did not succeed), then re-propagate.
+    let completion;
+    try {
+        completion = await completeWorkflowStep({
+            runId,
+            stepId,
+            status: "completed",
+            summary,
+            evidence: result.evidence,
+            summaryJudge,
+            ...lease,
+        });
+    }
+    catch (err) {
+        if (gateUnit)
+            await journalGateEvaluationFinish(gateUnit, true, undefined);
+        throw err;
+    }
+    const rejection = "ok" in completion && completion.ok === false ? completion : undefined;
+    if (gateUnit) {
+        await journalGateEvaluationFinish(gateUnit, judgeState.errored, rejection);
+    }
+    if (!rejection) {
+        return { kind: "advanced", ...(summaryOverride !== undefined ? { summaryOverride } : {}) };
+    }
+    if (loopsRemaining) {
+        return { kind: "retry", gateFeedback: { feedback: rejection.feedback, missing: rejection.missing } };
+    }
+    return {
+        kind: "gate-exhausted",
+        gateRejection: { stepId, missing: rejection.missing, feedback: rejection.feedback },
+    };
+}
+// ── Small helpers ────────────────────────────────────────────────────────────
+function safeJson(value) {
+    try {
+        return JSON.stringify(value) ?? "null";
+    }
+    catch {
+        return "null";
+    }
+}
+function clip(text, max) {
+    return text.length > max ? `${text.slice(0, max)}…` : text;
+}